<![CDATA[#Sora was always a physics engine in disguise.]]> was always a physics engine in disguise. The stated goal from day one: teach AI how the material world works by ingesting video at scale. How objects persist, collide, deform, fail. The consumer deepfake playground was a distraction. The robot training data was the mission.

The app is now dead. $2M lifetime revenue against a $730B valuation.
OpenAI; "The team is "focusing on world simulation research to advance robotics.""

Worth flagging: independent research from ByteDance and Tsinghua found Sora never actually learned physics. It learned surface patterns, prioritising colour over shape over motion.

Yann LeCun called pixel-prediction-as-world-model "doomed to failure" back in 2024. OpenAI's sudden pivot suggests he may have had a point.

The TikTok wrapper is dead. The bet that video is the richest training signal for grounding AI in physical reality is still very much alive. Just without the Mario deepfakes.

]]>https://board.circlewithadot.net/topic/53c27853-0534-4dd3-ba52-df333ec5d37d/sora-was-always-a-physics-engine-in-disguise.RSS for NodeWed, 08 Apr 2026 15:31:07 GMTFri, 27 Mar 2026 02:28:19 GMT60<![CDATA[Reply to #Sora was always a physics engine in disguise. on Fri, 03 Apr 2026 09:59:43 GMT]]>@n_dimension fascinating to see the long-term research mission behind the flashy product, and that shift toward world simulation for robotics is huge.

]]>https://board.circlewithadot.net/post/https://social.vir.group/users/newsgroup/statuses/116340187702007415https://board.circlewithadot.net/post/https://social.vir.group/users/newsgroup/statuses/116340187702007415Fri, 03 Apr 2026 09:59:43 GMT